Research On Fast Noise Reduction Technology Of Point Cloud Data Based On Multi-core DSP

The processing techniques of 3D point cloud data have great significance in the research fields of lidar,deep learning,computer vision,motion-sensing game,unmanned vehicle,reverse engineering and so on.In the process of 3D point cloud data acquisition,the data are inevitably polluted by the different levels of noise.The noises in the 3D point cloud degrade severely the following-up point cloud processing tasks.Therefore,noise reduction is very important in the processing techniques of 3D point cloud data.Traditional point cloud processing is based on the PC platform.Because of its bulky and high power consumption,it is not suitable for engineering applications.Multi-core DSP has the characteristics of small size,low power consumption,and high processing speed,which make it more suitable for engineering applications.So it is very meaningful to study the noise reduction technology of point cloud data based on multi-core DSP for miniaturization and integration.The noises in the point cloud data are complex.According to the noises distribution characteristics in the point cloud,this paper divides it into two categories: the first category is called the drift noises point,which is expressed as the point of the noises that are far away from the body of the point cloud and are sparse and scattered over it;the second category is called the mixed noise point,which is expressed as the point of the noises that are mixed with the main body of the point cloud.According to the characteristics of noise in point cloud data,this paper does the following research:First of all,for the drift noise points,the denoising algorithm is mainly used to remove them.The non-iterative dual threshold denoising algorithm is proposed in this paper.Experiments are conducted to compare the proposed method with the K-means cluster denoising method,the improved K-means cluster denoising method based on the K-d tree and the local density denoising method.The results show that compared with the K-means cluster denoising method,the improved K-means cluster denoising method based on the K-d tree and the local density denoising method,the computing time required by the presented method is reduced by 38%,18%,and 84% respectively.The denoising accuracy is better than 84% at 40% noisy environment.Secondly,aiming at the mixed noise points,the smoothing algorithm is mainly used to smoothing them.Experiments are conducted to compare the bilateral filter smoothing method with the guide filter smoothing method.The results show that the guide filter smoothing method is better than the bilateral filter smoothing method in terms of mean square error,SNR and taking time.Especially,compared with the bilateral filter smoothing method,the computing time required by the guide filter smoothing method is reduced by 38%.Then,the proposed denoising algorithm is designed in parallel.Aiming at the problem that the guiding filter smoothing algorithm can not be processed in parallel,a non-uniform guiding filter smoothing algorithm is proposed in this paper.Finally,the noise reduction of the point cloud is carried out on the multi-core DSP.The hardware results are compared with the simulation results of PC,the rationality and effectiveness of parallel point cloud noise reduction are demonstrated.